AI News — October 25, 2026: Cerebras Spans 59 Wafers for 3T Kimi, Five Qwens Tie Claude Fable 5

Cerebras published a technical writeup detailing how it serves the 3-trillion-parameter Kimi K3 across 59 WSE-3T wafers using pipeline parallelism, 4-bit weight quantization and an 8-bit KV cache, reaching 1,000–2,000+ tokens per second at roughly 1.18 MW total draw and about 4.1 joules per token. Separately, a GitHub project called GVS5H orchestrates five Qwen3.8-27B instances through a shared filesystem and hits 92.4% pass@1 on LiveCodeBench hard problems, essentially tied with Claude Fable 5 at 90.4%, though commenters flagged the agentic-loop comparison as unfair on a saturated benchmark. Apollo chief economist Torsten Slok found no measurable AI gains in revenue per worker, margins or total factor productivity across the S&P 493, and Apricitas documented a four-year decline in US media, film and arts employment, with Hollywood alone losing about a third of its workforce.

Good morning. The China capability story keeps compounding — Kimi K3.1 Pro landed yesterday, and today Cerebras is showing how it serves a 3-trillion-parameter Kimi model across 59 wafers while a scrappy harness runs five Qwen instances to match a frontier lab. Meanwhile, Apollo’s chief economist says there’s still no measurable AI payoff outside tech, and the creative sector is 200,000 jobs lighter than it was four years ago. Cerebras deploys trillion-parameter Kimi on 59 wafers. Cerebras published a technical writeup https://www.cerebras.ai/blog/how-we-deploy-trillion-parameter-frontier-models-on-cerebras of how it serves the 3-trillion-parameter Kimi K3 using pipeline parallelism across 59 WSE-3T wafers, 4-bit weight quantization, and an 8-bit KV cache, hitting 1,000–2,000+ tokens per second. One HN commenter did the math on the whole rig: roughly 1.18 MW total draw at ~4.1 joules per token. Others pointed out the gap between the marketing and reality — publicly accessible models on Cerebras are still limited to things like GPT OSS 120B and Qwen 3.8 27B, which users describe as fast but often wrong. Five Qwens in a trenchcoat match Fable 5. A GitHub project called GVS5H https://github.com/slee-persis/GVS5H orchestrates five Qwen3.8-27B instances through a shared filesystem — plan.md, tasks.json, a verifier role — and hits 92.4% pass@1 on LiveCodeBench hard problems, essentially tied with Claude Fable 5 at 90.4%. The catch, which several HN commenters flagged immediately: giving models an agentic loop on a benchmark designed for single-pass evaluation isn’t really a fair comparison, and LiveCodeBench is a saturated 2025-era benchmark with no numbers here against Astra, Sol, or Opus 5.5. Still, the cost story is real if you’re running locally on hardware you already own. No sign of AI payoff outside the tech sector. Apollo chief economist Torsten Slok finds no measurable gains https://www.apollo.com/wealth/insights-news/insights/daily-spark/no-signs-of-ai-payoff-outside-the-tech-sector in revenue per worker, margins, or total factor productivity across the S&P 493, and argues most of corporate America now treats token spending as a cost center. The HN thread offered two competing explanations: the electrification-lag theory non-software industries need capital reorganization that takes a decade , and the vendor-capture theory, nicely summarized by one commenter — “approximately all of the gain is captured as revenue for AI companies.” Another pointed out that software firms can deploy non-deterministic systems because they have test suites as validation gates; most industries don’t, so they can’t tell when the model is bullshitting them. 200,000 creative-sector jobs gone, AI’s role unclear. Apricitas documents https://www.apricitas.io/p/ai-and-the-fall-of-the-creative-class a four-year decline in US media, film, and arts employment comparable only to the 2001 and 2008 recessions, with Hollywood alone losing about a third of its workforce. The piece is careful about causation — streaming economics, interest rates, consolidation, and offshoring all contribute — but notes that digital creative work is shrinking while in-person entertainment grows, which fits the AI story. The clearest casualty, per one HN commenter, is commercial illustration: the timeline tracks closely with Stable Diffusion’s arrival. Another pointed to a confounder the article skipped entirely — the influencer economy, which reportedly pays out $25–50B a year to creators. Block’s Dessa voice model goes full duplex. Jack Dorsey’s Block released Dessa https://techcrunch.com/2026/10/24/blocks-new-voice-model-sounds-eerily-human/ , a speech-to-speech model trained end-to-end on human speech without text as an intermediary, running locally on a single NVIDIA GPU rack. Block claims emotion is baked into the model rather than steered by voice tags, producing laughter, fatigue, and near-tearful cadences natively. Some researchers quoted in the piece are calling it a “GPT-3 moment for speech,” though whether there’s still hidden tagging under the hood is an open question. Pennsylvania mandates AI ad labels. Governor Josh Shapiro signed the AI Deceptive Advertising Act, The Verge reports https://www.theverge.com/ai-artificial-intelligence/1018187/pennsylvania-ai-deceptive-advertising-act , requiring disclosure on AI-generated or “significantly modified” ads across image, video, audio, and text — with reach into other states if the ad also runs in Pennsylvania. The loopholes are wide: “significantly modified” isn’t defined, satire and “evidently fictional” content are exempt, and the standard for “false consumer perception” is vague. Grok picks R, GPT picks D. A Convovo/Cornell study found that when forced via API to recommend candidates across ~500 races, Grok 4.7 picked Republicans 92% of the time and GPT-6.1 Sol picked Democrats 97% of the time, agreeing on fewer than one in ten races, TechCrunch reports https://techcrunch.com/2026/10/24/forced-to-pick-groks-ai-model-chooses-republicans-openais-picks-democrats/ . OpenAI noted that consumer ChatGPT declines these questions and that API behavior isn’t typical user experience. The researchers’ real worry isn’t direct endorsements but the softer bias in how models frame issues and candidates in ordinary conversation. A 360M model that pretends to be a terminal. A developer fine-tuned Qwen2.5-Coder-0.5B to simulate a terminal purely through next-token prediction — no actual shell underneath — and got it convincingly running vim, tmux, and snake https://simedw.com/2026/10/21/world-model-terminal/ by treating the screen as a character grid and learning from Docker-recorded sessions. HN comparisons ranged from the classic “pretend to be a terminal” ChatGPT prompt to Anthropic’s use of similar world-model simulations for alignment evaluations. Not useful, as the author admits, but a clean thought experiment on where prediction stops and computation begins. The AI graveyard grows. TechCrunch is maintaining a running list https://techcrunch.com/2026/10/24/the-ai-graveyard-a-running-list-of-projects-and-startups-that-didnt-make-it/ of dead AI products, with Character.ai’s bankruptcy and workflow tool Relay among the recent entries. OpenAI shows up repeatedly too — ChatGPT Atlas, Operator, standalone DALL-E, and the “super app” redesign have all been shelved. S&P Global’s data suggests about 42% of corporate AI initiatives are eventually abandoned, which puts the overall mood of the Apollo report in some context. That’s the morning. The pattern holds: open-weight models from China keep catching up, orchestration tricks keep squeezing more out of smaller models, and the economic case outside tech keeps not showing up in the numbers.